Deconstructing the AI Landscape

Seriously, another AI post? Right, another one. But for a reason. Like Andrew Ng, a well known AI thought leader already said “AI is the new electricity”. And yes, I agree: Similar to electricity, AI seems to be everywhere. Based on recent reports, the ratio of startups embracing AI grew from only one in 50 in 2013 to one in 12 as of 2019. Additionally, with ever growing tech complexity and variety of manifestations, it becomes increasingly difficult to distinguish all the “AI companies” out there. Exponential growth and omnipresence together with increasing complexity and decreasing understanding inevitably leads to scepticism and confusion.

Source: http://i.imgur.com/YIS9f.jpg

Although there are hundreds and thousands of frameworks, landscapes and cheat sheets available, I had variety of discussions which led me to conclude that dependent on our diverse backgrounds, we all have something different in mind when referring to AI. Some might think of the cognitive areas or the hierarchy of needs, whereas others tend to think in boxes of narrow, general and super AI or just any system that improves itself over time. Based on the variety of perceptions, people frequently compare oranges and apples and talk at cross purposes. There seems to be no common framework which easily allows us to classify AI startups, ask the right questions and ultimately make the right decisions. And nothing but this is the purpose of this post: to shed some light on today’s chaotic AI world by sharing my framework together with a set of simple but powerful questions.

Intelligent or not?

The fact that AI is such a broad concept forces us to agree on the lowest common denominator which summarises AI as “a set of intelligent technologies rather than an approach of mimicking human intelligence”. This raises the question of what we actually understand as “intelligent”. To answer this question, I’d like to build on the three types of analytics — descriptive, predictive and prescriptive. While descriptive solutions are data-centric, retrospective and contain no intelligence at all, predictive and prescriptive solutions are action-centric, future-oriented and contain some level of intelligence per definition. In short, the latter two are the ones which we consider as “intelligent technologies” or AI.

The AI crunch question

Once agreed on what is AI and what isn’t, we should ask ourselves the one and only AI crunch question to move ahead:

Do you want to build a core-AI tech component which is applicable to variety of use cases and independent from a specific industry (horizontal), or do you aim to solve a very use-case- or domain-specific issue by applying AI (vertical)?

Hereby, we end up splitting the pool of intelligent companies into horizontal “tech-providers” and vertical “integrators”. We classify the company as a tech provider in case their product is applicable to a variety of industries and use cases. If their business is focused (or even limited) to one specific industry or use case, however, we classify it as an integrator. While transitioning gradually from horizontal to vertical or vice versa is rather common and natural, being in both boxes at the same time from day one is not. For me, the latter indicates an inconsistent vision and a lack of focus. In summary, the crunch question helps us to compare apples to apples and understand the overall vision of a company.

Horizontal AI — The “Tech-Providers”

For the group of horizontal tech-providers who independently of use cases or industries aim to build one or more specific components better than anyone else, we get to the next level of granularity by distributing them across the data value chain. Again, there are different structures available but the lowest common denominator seems to be the following 5-step process: Collect (data), Process (data), Explore (data), Build (models) and Deploy (models). We may split this process into even more granular sub-steps as displayed above. A detailed description of every sub-step would clearly go beyond the scope of this framework-focused post, so I will keep this for another post and provide a high-level overview below (please excuse fuzzy explanations here).

Collect data from relevant sources
Internal data: primary data which is generated within an institution such as through sensors, ERP systems or written notes
External data: secondary data which is collected from external parties such as data providers or public databases

Process data in a secure, structured and efficient way
Infrastructure: enable structured funnelling via cloud or on-premise
Privacy: allow secure and regulation-compliant data sharing
Transformation: bring data into suitable format for exploration

Explore the characteristics of the dataset
Cleaning: handle missing values, wrong data types and other data issues
Labeling: add meaningful tags to observations (dependent variable)
Feature engineering: combine features (independent variables)

Build variety of models and identify the best performing one
Architecture: environments and workbenches such as Github or TensorFlow
Learning: manual or automated model building
Validating: identification of most suitable model

Deploy selected models for dynamic improvement
Direct: integrate model as code-snipped into environment or device
Indirect: integrate model through 3rd party platforms or APIs

Although, most tech-providers focus majority of their resources on one very specific sub-step, they are likely to add adjacent sub-steps over time to increase their overall value proposition. For example, an Automated Machine Learning company would clearly fall into the Learning sub-step of the Build step because its core value proposition is to build and benchmark a variety of ML models with ease. However, the company will likely start adding adjacent sub-steps such as Cleaning, Feature Engineering and Validating to provide a better overall solution. Whether they make these additional sub-steps in-house or buy them externally depends on the closeness of the adjacent sub-steps to their core tech.

Horizontal AI — A Subset of EnterpriseTech & DevTech

Depending on whether the product is used by enterprises in general or by developers more specifically, we classify these horizontal players either as EnterpriseTech or DevTech companies. Essentially, horizontal AI companies are a subset of EnterpriseTech and/or DevTech. Some of the most crucial factors for these companies are listed below:

• Focus, focus, focus! While your tech can be applied to variety of industries and use-cases, we recommend finding the ones where you can add most value (either through systematic analysis and ranking or incrementally through opportunistic PoCs) and focus accordingly. Gradually expand (move ahead) once you dominated (or lost) them.
• Don’t lose yourself in one industry! Establish a scalable go-to-market strategy which is replicable across industries, e.g., build partnerships and incentivise (e.g., through revenue share model) 3rd parties to distribute your product rather than owning multiple industry-specific distribution channels yourself. This provides room to iterate across industries and utilise existing infrastructures.
• Be aware of your resources! “Zero to One” taught us about iterating your product based on your customer’s demands. Focus your resources on those requests/needs which are applicable across customers and industries. Don’t customise too early, it might be a one-way rabbit hole.
• Most horizontal AI building blocks are likely to be commoditised in the future — establish sustainable differentiation and defensibility mechanisms through backward- and forward-looking lock-in effects.
• Be close to research, build a strong tech team and proprietary IP.
• Early traction does not matter as much compared to vertical players. Data on usage behaviour (MAU/DAU, uptime/downtime, churn, etc.) is essential though.
• Productise! Move from PoC service business to product-based licensing.
• …

Vertical AI — The “Integrators”

For the group of vertical integrators who aim to solve one very use-case- or industry-specific issue, we categorise them across the major verticals. Although most of these industries face fundamentally different issues, those integrators building AI-enabled solutions to solve them face very similar ones.

Make core and buy non-core

Obviously, the vertical framework is pretty straight-forward and while some of these players build parts of their AI in-house and from scratch, most of them decide to rely on available solutions (from specialised tech-providers) and stack them together to build their unique customised AI product. Unfortunately, it seems like if still too many companies waste majority of their resources as they seek to reinvent the wheel and build everything from scratch rather than staying on the shoulders of giants, i.e., those who specialise in it. From my perspective, companies should build strategically relevant components in-house and outsource the rest. Without focus, you will likely compete with horizontal tech-providers and other vertical integrators at the same time. Further crucial factors are listed below:

• Spin the “AI Flywheel”! Domain expertise and proprietary (use-case-specific) datasets help you to add value from day one: stack existing (and/or proprietary) models together, feed them with your data, add value to your first customers, collect more data, improve models, add more value … #DataNetworkEffects
• Define and execute a superior land-grab, pricing and data acquisition strategy. Often enough, early partners allow you to train your models with their proprietary data in return for free use of your models. Don’t scare them away with the wrong pricing (correct pricing is a topic for itself though..)
• Be close to your customers and build a strong domain expert team. Customer is king — know them in and out, i.e., their needs and processes.
• Early traction matters a lot compared to horizontal players.
• …

So what?

As discussed at the outset, the purpose of this post is to share our high-level framework which helps us to classify AI companies and identify the crucial points for follow-on deep dives more accurately. I established this framework over time and iterated it with feedback from founders, researchers, other investors and conference participants (thanks a lot). It is certainly valuable for us and hopefully, it helps you too.

Although the field of AI is maturing, we strongly believe there are tons of opportunities to build category-leading companies in the space. Are you a founder, industry expert, VC or researcher interested in the field of AI? I’d be more than happy to learn about your work, so feel free to reach out at andre@earlybird.com.

Liked this piece?
>>Follow me on twitter for more AI and VC related stuff.